Automated Classification of Resting-State fMRI ICA Components Using a Deep Siamese Network

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Manual classification of resting state networks (RSNs) from Independent Component Analysis (ICA) is time-consuming and requires specialized expertise.
• Automated methods are needed for efficient and reliable RSN classification in large-scale neuroimaging studies.

Purpose of the Study:

• To develop and evaluate a deep learning approach for automatic classification of single-subject ICA-derived RSNs.
• To assess the performance of the proposed method against traditional techniques and its robustness to variations.

Main Methods:

• A supervised deep learning framework utilizing a Siamese Network architecture was employed.
• The method learns discriminative feature representations for RSN classification, enabling one-shot learning.
• Performance was evaluated on holdout and external datasets, comparing against Convolutional Neural Networks (CNNs) and template matching.

Main Results:

• The Siamese Network approach achieved 100% accuracy on a holdout dataset and over 99% on an external dataset.
• The method demonstrated robustness to scan-rescan variability.
• Altered functional connectivity in default mode and salience networks was identified in mild and severe TBI groups compared to healthy controls.

Conclusions:

• The proposed deep learning method offers a highly accurate and efficient solution for automatic RSN classification.
• This technique facilitates the identification of functional brain network alterations in neurological conditions like TBI.
• The approach supports generalization to new RSN classes with minimal training data.